2018-09-04 18:49:07

by Laura Abbott

[permalink] [raw]
Subject: [PATCHv3] iscsi-target: Don't use stack buffer for scatterlist

Fedora got a bug report of a crash with iSCSI:

kernel BUG at include/linux/scatterlist.h:143!
...
RIP: 0010:iscsit_do_crypto_hash_buf+0x154/0x180 [iscsi_target_mod]
...
Call Trace:
? iscsi_target_tx_thread+0x200/0x200 [iscsi_target_mod]
iscsit_get_rx_pdu+0x4cd/0xa90 [iscsi_target_mod]
? native_sched_clock+0x3e/0xa0
? iscsi_target_tx_thread+0x200/0x200 [iscsi_target_mod]
iscsi_target_rx_thread+0x81/0xf0 [iscsi_target_mod]
kthread+0x120/0x140
? kthread_create_worker_on_cpu+0x70/0x70
ret_from_fork+0x3a/0x50

This is a BUG_ON for using a stack buffer with a scatterlist.
There are two cases that trigger this bug. Switch to using a
dynamically allocated buffer for one case and do not assign
a NULL buffer in another case.

Signed-off-by: Laura Abbott <[email protected]>
---
v3: Switched to kcalloc to drop an extra memset call. I was asked
offline about kmalloc vs. kmalloc_array and ended up asking about this
at the security summit since I was there last week. The preferred style
is kmalloc_array for all arrays, even those of one byte characters.
I will also switch it if people feel really strongly about this or I've
misunderstood the array guidelines.
---
drivers/target/iscsi/iscsi_target.c | 22 ++++++++++++++--------
1 file changed, 14 insertions(+), 8 deletions(-)

diff --git a/drivers/target/iscsi/iscsi_target.c b/drivers/target/iscsi/iscsi_target.c
index 94bad43c41ff..2a586b58a011 100644
--- a/drivers/target/iscsi/iscsi_target.c
+++ b/drivers/target/iscsi/iscsi_target.c
@@ -1416,7 +1416,8 @@ static void iscsit_do_crypto_hash_buf(struct ahash_request *hash,

sg_init_table(sg, ARRAY_SIZE(sg));
sg_set_buf(sg, buf, payload_length);
- sg_set_buf(sg + 1, pad_bytes, padding);
+ if (padding)
+ sg_set_buf(sg + 1, pad_bytes, padding);

ahash_request_set_crypt(hash, sg, data_crc, payload_length + padding);

@@ -3910,10 +3911,14 @@ static bool iscsi_target_check_conn_state(struct iscsi_conn *conn)
static void iscsit_get_rx_pdu(struct iscsi_conn *conn)
{
int ret;
- u8 buffer[ISCSI_HDR_LEN], opcode;
+ u8 *buffer, opcode;
u32 checksum = 0, digest = 0;
struct kvec iov;

+ buffer = kcalloc(ISCSI_HDR_LEN, sizeof(*buffer), GFP_KERNEL);
+ if (!buffer)
+ return;
+
while (!kthread_should_stop()) {
/*
* Ensure that both TX and RX per connection kthreads
@@ -3921,7 +3926,6 @@ static void iscsit_get_rx_pdu(struct iscsi_conn *conn)
*/
iscsit_thread_check_cpumask(conn, current, 0);

- memset(buffer, 0, ISCSI_HDR_LEN);
memset(&iov, 0, sizeof(struct kvec));

iov.iov_base = buffer;
@@ -3930,7 +3934,7 @@ static void iscsit_get_rx_pdu(struct iscsi_conn *conn)
ret = rx_data(conn, &iov, 1, ISCSI_HDR_LEN);
if (ret != ISCSI_HDR_LEN) {
iscsit_rx_thread_wait_for_tcp(conn);
- return;
+ break;
}

if (conn->conn_ops->HeaderDigest) {
@@ -3940,7 +3944,7 @@ static void iscsit_get_rx_pdu(struct iscsi_conn *conn)
ret = rx_data(conn, &iov, 1, ISCSI_CRC_LEN);
if (ret != ISCSI_CRC_LEN) {
iscsit_rx_thread_wait_for_tcp(conn);
- return;
+ break;
}

iscsit_do_crypto_hash_buf(conn->conn_rx_hash, buffer,
@@ -3964,7 +3968,7 @@ static void iscsit_get_rx_pdu(struct iscsi_conn *conn)
}

if (conn->conn_state == TARG_CONN_STATE_IN_LOGOUT)
- return;
+ break;

opcode = buffer[0] & ISCSI_OPCODE_MASK;

@@ -3975,13 +3979,15 @@ static void iscsit_get_rx_pdu(struct iscsi_conn *conn)
" while in Discovery Session, rejecting.\n", opcode);
iscsit_add_reject(conn, ISCSI_REASON_PROTOCOL_ERROR,
buffer);
- return;
+ break;
}

ret = iscsi_target_rx_opcode(conn, buffer);
if (ret < 0)
- return;
+ break;
}
+
+ kfree(buffer);
}

int iscsi_target_rx_thread(void *arg)
--
2.17.1



2018-09-04 19:05:24

by Mike Christie

[permalink] [raw]
Subject: Re: [PATCHv3] iscsi-target: Don't use stack buffer for scatterlist

On 09/04/2018 01:47 PM, Laura Abbott wrote:
> Fedora got a bug report of a crash with iSCSI:
>
> kernel BUG at include/linux/scatterlist.h:143!
> ...
> RIP: 0010:iscsit_do_crypto_hash_buf+0x154/0x180 [iscsi_target_mod]
> ...
> Call Trace:
> ? iscsi_target_tx_thread+0x200/0x200 [iscsi_target_mod]
> iscsit_get_rx_pdu+0x4cd/0xa90 [iscsi_target_mod]
> ? native_sched_clock+0x3e/0xa0
> ? iscsi_target_tx_thread+0x200/0x200 [iscsi_target_mod]
> iscsi_target_rx_thread+0x81/0xf0 [iscsi_target_mod]
> kthread+0x120/0x140
> ? kthread_create_worker_on_cpu+0x70/0x70
> ret_from_fork+0x3a/0x50
>
> This is a BUG_ON for using a stack buffer with a scatterlist.
> There are two cases that trigger this bug. Switch to using a
> dynamically allocated buffer for one case and do not assign
> a NULL buffer in another case.
>
> Signed-off-by: Laura Abbott <[email protected]>
> ---
> v3: Switched to kcalloc to drop an extra memset call. I was asked
> offline about kmalloc vs. kmalloc_array and ended up asking about this
> at the security summit since I was there last week. The preferred style
> is kmalloc_array for all arrays, even those of one byte characters.
> I will also switch it if people feel really strongly about this or I've
> misunderstood the array guidelines.
> ---
> drivers/target/iscsi/iscsi_target.c | 22 ++++++++++++++--------
> 1 file changed, 14 insertions(+), 8 deletions(-)
>
> diff --git a/drivers/target/iscsi/iscsi_target.c b/drivers/target/iscsi/iscsi_target.c
> index 94bad43c41ff..2a586b58a011 100644
> --- a/drivers/target/iscsi/iscsi_target.c
> +++ b/drivers/target/iscsi/iscsi_target.c
> @@ -1416,7 +1416,8 @@ static void iscsit_do_crypto_hash_buf(struct ahash_request *hash,
>
> sg_init_table(sg, ARRAY_SIZE(sg));
> sg_set_buf(sg, buf, payload_length);
> - sg_set_buf(sg + 1, pad_bytes, padding);
> + if (padding)
> + sg_set_buf(sg + 1, pad_bytes, padding);
>
> ahash_request_set_crypt(hash, sg, data_crc, payload_length + padding);
>
> @@ -3910,10 +3911,14 @@ static bool iscsi_target_check_conn_state(struct iscsi_conn *conn)
> static void iscsit_get_rx_pdu(struct iscsi_conn *conn)
> {
> int ret;
> - u8 buffer[ISCSI_HDR_LEN], opcode;
> + u8 *buffer, opcode;
> u32 checksum = 0, digest = 0;
> struct kvec iov;
>
> + buffer = kcalloc(ISCSI_HDR_LEN, sizeof(*buffer), GFP_KERNEL);
> + if (!buffer)
> + return;
> +
> while (!kthread_should_stop()) {
> /*
> * Ensure that both TX and RX per connection kthreads
> @@ -3921,7 +3926,6 @@ static void iscsit_get_rx_pdu(struct iscsi_conn *conn)
> */
> iscsit_thread_check_cpumask(conn, current, 0);
>
> - memset(buffer, 0, ISCSI_HDR_LEN);
> memset(&iov, 0, sizeof(struct kvec));
>
> iov.iov_base = buffer;
> @@ -3930,7 +3934,7 @@ static void iscsit_get_rx_pdu(struct iscsi_conn *conn)
> ret = rx_data(conn, &iov, 1, ISCSI_HDR_LEN);
> if (ret != ISCSI_HDR_LEN) {
> iscsit_rx_thread_wait_for_tcp(conn);
> - return;
> + break;
> }
>
> if (conn->conn_ops->HeaderDigest) {
> @@ -3940,7 +3944,7 @@ static void iscsit_get_rx_pdu(struct iscsi_conn *conn)
> ret = rx_data(conn, &iov, 1, ISCSI_CRC_LEN);
> if (ret != ISCSI_CRC_LEN) {
> iscsit_rx_thread_wait_for_tcp(conn);
> - return;
> + break;
> }
>
> iscsit_do_crypto_hash_buf(conn->conn_rx_hash, buffer,
> @@ -3964,7 +3968,7 @@ static void iscsit_get_rx_pdu(struct iscsi_conn *conn)
> }
>
> if (conn->conn_state == TARG_CONN_STATE_IN_LOGOUT)
> - return;
> + break;
>
> opcode = buffer[0] & ISCSI_OPCODE_MASK;
>
> @@ -3975,13 +3979,15 @@ static void iscsit_get_rx_pdu(struct iscsi_conn *conn)
> " while in Discovery Session, rejecting.\n", opcode);
> iscsit_add_reject(conn, ISCSI_REASON_PROTOCOL_ERROR,
> buffer);
> - return;
> + break;
> }
>
> ret = iscsi_target_rx_opcode(conn, buffer);
> if (ret < 0)
> - return;
> + break;
> }
> +
> + kfree(buffer);
> }
>
> int iscsi_target_rx_thread(void *arg)
>

Ok.

Reviewed-by: Mike Christie <[email protected]>

2018-09-12 00:47:40

by Martin K. Petersen

[permalink] [raw]
Subject: Re: [PATCHv3] iscsi-target: Don't use stack buffer for scatterlist


Laura,

> There are two cases that trigger this bug. Switch to using a
> dynamically allocated buffer for one case and do not assign
> a NULL buffer in another case.

Applied to 4.20/scsi-queue, thank you!

--
Martin K. Petersen Oracle Linux Engineering

2018-09-12 00:49:10

by Martin K. Petersen

[permalink] [raw]
Subject: Re: [PATCHv3] iscsi-target: Don't use stack buffer for scatterlist


> Applied to 4.20/scsi-queue, thank you!

4.19/scsi-fixes, that is...

--
Martin K. Petersen Oracle Linux Engineering